Incident 62: Bad AI-Written Christmas Carols

Description: Janelle Shane, an AI research scientist, used 240 popular Christmas carols to train a neural network to write its own carols
Alleged: Janelle Shane developed and deployed an AI system, which harmed Carollers.

Suggested citation format

Yampolskiy, Roman. (2017-12-23) Incident Number 62. in McGregor, S. (ed.) Artificial Intelligence Incident Database. Responsible AI Collaborative.

Incident Stats

Incident ID
62
Report Count
2
Incident Date
2017-12-23
Editors
Sean McGregor

Tools

New ReportNew ReportDiscoverDiscover

CSET Taxonomy Classifications

Taxonomy Details

Full Description

Janelle Shane, an AI research scientist, used 240 popular Christmas carols to train a neural network to write its own carols. The output closely mimicked a Christmas carol, but made no logical sense.

Short Description

Janelle Shane, an AI research scientist, used 240 popular Christmas carols to train a neural network to write its own carols

Severity

Unclear/unknown

Harm Distribution Basis

Age

AI System Description

Neural network designed by Janelle Shane meant to write Christmas carols.

System Developer

Janelle Shane

Sector of Deployment

Arts, entertainment and recreation

Relevant AI functions

Cognition, Action

AI Techniques

Neural network

AI Applications

language processing, language output

Named Entities

Janelle Shane

Technology Purveyor

Janelle Shane

Beginning Date

2017-12-22T08:00:00.000Z

Ending Date

2017-12-22T08:00:00.000Z

Near Miss

Unclear/unknown

Intent

Unclear

Lives Lost

No

Data Inputs

240 popular Christmas carols

Incidents Reports

Neural networks are a type of computer program that imitate the way that brains learn to solve problems. They’re used for face recognition, self-driving cars, language translation, financial decisions, and more. I mainly use them to write humor.

My process starts with a dataset - something that the neural network has to figure out how to imitate. Rather unfairly, I give it no instructions about whether it’s trying to write knock-knock jokes or invent Halloween costumes or begin a novel. It doesn’t even have any built-in knowledge of English. It definitely doesn’t have any clue what Christmas carols are, which is why this week’s experiment was so much fun.

The Times of London teamed up with reader/neural net hobbyist Erik Svensson to collect a mix of ancient and modern carols, about 240 carols in all, from “What Child is This?” to “Grandma Got Run Over by a Reindeer”.

When the neural network begins learning, it starts with a set of random rules about how to put one letter after another to make a Christmas carol. Since they are random, they are terrible rules, and when the neural network tries to apply them, it gets junk:

a a a na snn aca naa i s a i aa a a n a uugna nn na i a uaa a a i a a a a a sna aagt o aa

But it can check its rules against the real songs in its database and then make little tweaks to them that make them work slightly better. It learns to capitalize the first word of a line, it learns how often to do line breaks… and soon it begins to learn actual words - the most common ones first.

Hart fon the be the he br wong on the stor Christmas br he, or the wangChrist, Christ, on bn a me the stordHont on thr st bong the wor

I he a s de poog the stow tome on be ser snur

After the neural network has spent many more rounds refining its rules, it begins to look a lot like Christmas.

Now, neural networks tend to pick up tone and vocabulary pretty easily, but struggle with making sense. So, the carol-trained neural network learned to produce a lot of lines that sound - well, joyful, at least.

The cattle around the Christmas will be
A very special Christmas with me

Hurry Christmas to you

Cup on the earth!

Still the loudly candlelight
Would praise His name.

The babe, the Son of Mary.
He sumbled their flowers and all.

The lord of the glory dawnsGive us the leave all awayA star is spent and redShake a cup a strend from the skyChristmas is coming, the wind is come to you

Walkin’ him love, Dingle bells, jingle bells, jingle bells with bells are ringing

With a heart reindeerBut no more a stranger.

Santa baby, and Dancer, and Curry down

Happy HolidayWhen the snowflakes will call the world wakes to bringGlory bears and asses the air the angels sangAnd Christmas tree

(Curry is not in the input data. Nor is Dingle. The neural net likes to invent words that it thinks sound sort of carol-y)

Other carols it generated sounded a bit more morose.

The fire is sleeping.And crying,Love love

What a King

Let’s take the little children of the grave!

For some reason, the Sandman figures very prominently in the neural net’s Christmas mythology, despite having been mentioned in the dataset only once. Sometimes the neural net latches on to particular words for no reason I can see. Maybe it’s a Neil Gaiman fan.

The sandman so be joyful now it was born today!Gloria in excelsis Deo.

The sandman bright before Him.The holly bears a berry bears

And star in the snow is born today!

The sandman so love to seek the world

The sandman so love so deep and sing and the sun

And this? In retrospect, I should have seen this coming:

The world and joy of the sleigh
Santa baby bore sweet Jesus Christ

The holly bears a berry,
And all the reindeer of the sky

The holly bears a berry and reindeer
He was born today!

And Santa baby bore sweet Jesus Christ,And the chimney the angels sing.

When the snowman brings the snowChristmas treeLet’s take that road beforeAnd Santa Claus comes tonightHe will bring us goodness and light

Santa baby, a blitzen,And he was the sun and reindeer and earth.

The Saviour of the chimney tonight

The story of the chimney seeSanta baby, and blood and joyous so world and joy and good will to seeSanta baby bore sweet Jesus Christ

Fa la la la la la la, la la la la la la la la.

King of toys and hippopotamuses [sic] full of the light of that stood at the dear Son of Santa Claus
He was born in a wonderful christmas tree

Run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolf the new born King.

Christmas Carols, generated by a neural network

Elon Musk is terrified that artificial intelligence might spell the end of humanity.

Thankfully, it looks like we're not quite there yet. Colorado-based research scientist Janelle Shane has trained a neural network (a type of machine-learning algorithm) to write its own Christmas carols, and the results are...interesting.

SEE ALSO: AI could discover the next best esports pro player

Shane trained the algorithm to imitate a set of 240 popular Christmas carols aggregated by the Times of London. The AI trained itself by continuously attempting to write carols, checking their accuracy against the carols in the dataset, and modifying its process accordingly.

Here's an excerpt from one, which Shane posted on her blog:

The story of the chimney see

Santa baby, and blood and joyous so world and joy and good will to see

Santa baby bore sweet Jesus Christ

Fa la la la la la la, la la la la la la la la.

King of toys and hippopotamuses [sic] full of the light of that stood at the dear Son of Santa Claus

He was born in a wonderful christmas tree

Run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolf the new born King.

As you can see, the resulting songs look a heck of a lot like Christmas carols, but they don't make a ton of sense.

It's actually somewhat impressive how well Shane's AI imitates the structure and syntax of a Christmas song. But it's clearly got a long way to go.

AI alarmists may want to holster their pitchforks for now. The robot overlords may have conquered strategy games, but when it comes to Christmas spirit, humans are still on top.

AI still sucks at writing Christmas Carols

Similar Incidents

By textual similarity

Did our AI mess up? Flag the unrelated incidents

TayBot

· 26 reports